Average treatment effect (ATE)

The average treatment effect is the mean difference between the outcome each unit would have had under treatment and the outcome it would have had under control, averaged across the whole population studied. In a randomized experiment, it is estimated simply as the difference between the treatment group's average outcome and the control group's average outcome, since randomization makes the two groups comparable counterfactuals for each other.

The ATE summarizes the impact for a "typical" unit, but it can mask a lot: a treatment that helps half the population and hurts the other half by the same amount produces an ATE of zero, indistinguishable on paper from a treatment that does nothing to anyone. This is the gap the ATE leaves for heterogeneous treatment effects analysis to fill, by asking whether the effect differs across identifiable subgroups rather than assuming it is uniform.

The ATE is the default number reported from most A/B tests and quasi-experimental studies because it answers the most common business question, did this work on average, but relying on it alone can hide both winners and losers within the tested population, which matters whenever a rollout decision could instead be targeted rather than all-or-nothing.